path: root/gfx/wr/swgl/src/composite.h

/* This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this
 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */

// Converts a pixel from a source format to a destination format. By default,
// just return the value unchanged as for a simple copy.
template <typename P, typename U>
static ALWAYS_INLINE P convert_pixel(U src) {
  return src;
}

// R8 format maps to BGRA value 0,0,R,1. The byte order is endian independent,
// but the shifts unfortunately depend on endianness.
template <>
ALWAYS_INLINE uint32_t convert_pixel<uint32_t>(uint8_t src) {
#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
  return (uint32_t(src) << 16) | 0xFF000000;
#else
  return (uint32_t(src) << 8) | 0x000000FF;
#endif
}

// RG8 format maps to BGRA value 0,G,R,1.
template <>
ALWAYS_INLINE uint32_t convert_pixel<uint32_t>(uint16_t src) {
  uint32_t rg = src;
#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
  return ((rg & 0x00FF) << 16) | (rg & 0xFF00) | 0xFF000000;
#else
  return (rg & 0xFF00) | ((rg & 0x00FF) << 16) | 0x000000FF;
#endif
}

// RGBA8 format maps to R.
template <>
ALWAYS_INLINE uint8_t convert_pixel<uint8_t>(uint32_t src) {
#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
  return (src >> 16) & 0xFF;
#else
  return (src >> 8) & 0xFF;
#endif
}

// RGBA8 formats maps to R,G.
template <>
ALWAYS_INLINE uint16_t convert_pixel<uint16_t>(uint32_t src) {
#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
  return ((src >> 16) & 0x00FF) | (src & 0xFF00);
#else
  return (src & 0xFF00) | ((src >> 16) & 0x00FF);
#endif
}

// R8 format maps to R,0.
template <>
ALWAYS_INLINE uint16_t convert_pixel<uint16_t>(uint8_t src) {
#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
  return src;
#else
  return uint16_t(src) << 8;
#endif
}

// RG8 format maps to R.
template <>
ALWAYS_INLINE uint8_t convert_pixel<uint8_t>(uint16_t src) {
#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
  return src & 0xFF;
#else
  return src >> 8;
#endif
}

template <bool COMPOSITE, typename P>
static inline void copy_row(P* dst, const P* src, int span) {
  // No scaling, so just do a fast copy.
  memcpy(dst, src, span * sizeof(P));
}

template <>
void copy_row<true, uint32_t>(uint32_t* dst, const uint32_t* src, int span) {
  // No scaling, so just do a fast composite.
  auto* end = dst + span;
  while (dst + 4 <= end) {
    WideRGBA8 srcpx = unpack(unaligned_load<PackedRGBA8>(src));
    WideRGBA8 dstpx = unpack(unaligned_load<PackedRGBA8>(dst));
    PackedRGBA8 r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
    unaligned_store(dst, r);
    src += 4;
    dst += 4;
  }
  if (dst < end) {
    WideRGBA8 srcpx = unpack(partial_load_span<PackedRGBA8>(src, end - dst));
    WideRGBA8 dstpx = unpack(partial_load_span<PackedRGBA8>(dst, end - dst));
    auto r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
    partial_store_span(dst, r, end - dst);
  }
}

template <bool COMPOSITE, typename P, typename U>
static inline void scale_row(P* dst, int dstWidth, const U* src, int srcWidth,
                             int span, int frac) {
  // Do scaling with different source and dest widths.
  for (P* end = dst + span; dst < end; dst++) {
    *dst = convert_pixel<P>(*src);
    // Step source according to width ratio.
    for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
      src++;
    }
  }
}

template <>
void scale_row<true, uint32_t, uint32_t>(uint32_t* dst, int dstWidth,
                                         const uint32_t* src, int srcWidth,
                                         int span, int frac) {
  // Do scaling with different source and dest widths.
  // Gather source pixels four at a time for better packing.
  auto* end = dst + span;
  for (; dst + 4 <= end; dst += 4) {
    U32 srcn;
    srcn.x = *src;
    for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
      src++;
    }
    srcn.y = *src;
    for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
      src++;
    }
    srcn.z = *src;
    for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
      src++;
    }
    srcn.w = *src;
    for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
      src++;
    }
    WideRGBA8 srcpx = unpack(bit_cast<PackedRGBA8>(srcn));
    WideRGBA8 dstpx = unpack(unaligned_load<PackedRGBA8>(dst));
    PackedRGBA8 r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
    unaligned_store(dst, r);
  }
  if (dst < end) {
    // Process any remaining pixels. Try to gather as many pixels as possible
    // into a single source chunk for compositing.
    U32 srcn = {*src, 0, 0, 0};
    if (end - dst > 1) {
      for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
        src++;
      }
      srcn.y = *src;
      if (end - dst > 2) {
        for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
          src++;
        }
        srcn.z = *src;
      }
    }
    WideRGBA8 srcpx = unpack(bit_cast<PackedRGBA8>(srcn));
    WideRGBA8 dstpx = unpack(partial_load_span<PackedRGBA8>(dst, end - dst));
    auto r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
    partial_store_span(dst, r, end - dst);
  }
}

template <bool COMPOSITE = false>
static NO_INLINE void scale_blit(Texture& srctex, const IntRect& srcReq,
                                 Texture& dsttex, const IntRect& dstReq,
                                 bool invertY, const IntRect& clipRect) {
  assert(!COMPOSITE || (srctex.internal_format == GL_RGBA8 &&
                        dsttex.internal_format == GL_RGBA8));
  // Cache scaling ratios
  int srcWidth = srcReq.width();
  int srcHeight = srcReq.height();
  int dstWidth = dstReq.width();
  int dstHeight = dstReq.height();
  // Compute valid dest bounds
  IntRect dstBounds = dsttex.sample_bounds(dstReq).intersect(clipRect);
  // Compute valid source bounds
  IntRect srcBounds = srctex.sample_bounds(srcReq, invertY);
  // If srcReq is outside the source texture, we need to clip the sampling
  // bounds so that we never sample outside valid source bounds. Get texture
  // bounds relative to srcReq and scale to dest-space rounding inward, using
  // this rect to limit the dest bounds further.
  IntRect srcClip = srctex.bounds() - srcReq.origin();
  if (invertY) {
    srcClip.invert_y(srcReq.height());
  }
  srcClip.scale(srcWidth, srcHeight, dstWidth, dstHeight, true);
  dstBounds.intersect(srcClip);
  // Check if clipped sampling bounds are empty
  if (dstBounds.is_empty()) {
    return;
  }

  // Calculate source and dest pointers from clamped offsets
  int srcStride = srctex.stride();
  int destStride = dsttex.stride();
  char* dest = dsttex.sample_ptr(dstReq, dstBounds);
  // Clip the source bounds by the destination offset.
  int fracX = srcWidth * dstBounds.x0;
  int fracY = srcHeight * dstBounds.y0;
  srcBounds.x0 = max(fracX / dstWidth, srcBounds.x0);
  srcBounds.y0 = max(fracY / dstHeight, srcBounds.y0);
  fracX %= dstWidth;
  fracY %= dstHeight;
  char* src = srctex.sample_ptr(srcReq, srcBounds, invertY);
  // Inverted Y must step downward along source rows
  if (invertY) {
    srcStride = -srcStride;
  }
  int span = dstBounds.width();
  for (int rows = dstBounds.height(); rows > 0; rows--) {
    switch (srctex.bpp()) {
      case 1:
        switch (dsttex.bpp()) {
          case 2:
            scale_row<COMPOSITE>((uint16_t*)dest, dstWidth, (uint8_t*)src,
                                 srcWidth, span, fracX);
            break;
          case 4:
            scale_row<COMPOSITE>((uint32_t*)dest, dstWidth, (uint8_t*)src,
                                 srcWidth, span, fracX);
            break;
          default:
            if (srcWidth == dstWidth)
              copy_row<COMPOSITE>((uint8_t*)dest, (uint8_t*)src, span);
            else
              scale_row<COMPOSITE>((uint8_t*)dest, dstWidth, (uint8_t*)src,
                                   srcWidth, span, fracX);
            break;
        }
        break;
      case 2:
        switch (dsttex.bpp()) {
          case 1:
            scale_row<COMPOSITE>((uint8_t*)dest, dstWidth, (uint16_t*)src,
                                 srcWidth, span, fracX);
            break;
          case 4:
            scale_row<COMPOSITE>((uint32_t*)dest, dstWidth, (uint16_t*)src,
                                 srcWidth, span, fracX);
            break;
          default:
            if (srcWidth == dstWidth)
              copy_row<COMPOSITE>((uint16_t*)dest, (uint16_t*)src, span);
            else
              scale_row<COMPOSITE>((uint16_t*)dest, dstWidth, (uint16_t*)src,
                                   srcWidth, span, fracX);
            break;
        }
        break;
      case 4:
        switch (dsttex.bpp()) {
          case 1:
            scale_row<COMPOSITE>((uint8_t*)dest, dstWidth, (uint32_t*)src,
                                 srcWidth, span, fracX);
            break;
          case 2:
            scale_row<COMPOSITE>((uint16_t*)dest, dstWidth, (uint32_t*)src,
                                 srcWidth, span, fracX);
            break;
          default:
            if (srcWidth == dstWidth)
              copy_row<COMPOSITE>((uint32_t*)dest, (uint32_t*)src, span);
            else
              scale_row<COMPOSITE>((uint32_t*)dest, dstWidth, (uint32_t*)src,
                                   srcWidth, span, fracX);
            break;
        }
        break;
      default:
        assert(false);
        break;
    }
    dest += destStride;
    // Step source according to height ratio.
    for (fracY += srcHeight; fracY >= dstHeight; fracY -= dstHeight) {
      src += srcStride;
    }
  }
}

template <bool COMPOSITE>
static void linear_row_blit(uint32_t* dest, int span, const vec2_scalar& srcUV,
                            float srcDU, sampler2D sampler) {
  vec2 uv = init_interp(srcUV, vec2_scalar(srcDU, 0.0f));
  for (; span >= 4; span -= 4) {
    auto srcpx = textureLinearPackedRGBA8(sampler, ivec2(uv));
    unaligned_store(dest, srcpx);
    dest += 4;
    uv.x += 4 * srcDU;
  }
  if (span > 0) {
    auto srcpx = textureLinearPackedRGBA8(sampler, ivec2(uv));
    partial_store_span(dest, srcpx, span);
  }
}

template <>
void linear_row_blit<true>(uint32_t* dest, int span, const vec2_scalar& srcUV,
                           float srcDU, sampler2D sampler) {
  vec2 uv = init_interp(srcUV, vec2_scalar(srcDU, 0.0f));
  for (; span >= 4; span -= 4) {
    WideRGBA8 srcpx = textureLinearUnpackedRGBA8(sampler, ivec2(uv));
    WideRGBA8 dstpx = unpack(unaligned_load<PackedRGBA8>(dest));
    PackedRGBA8 r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
    unaligned_store(dest, r);

    dest += 4;
    uv.x += 4 * srcDU;
  }
  if (span > 0) {
    WideRGBA8 srcpx = textureLinearUnpackedRGBA8(sampler, ivec2(uv));
    WideRGBA8 dstpx = unpack(partial_load_span<PackedRGBA8>(dest, span));
    PackedRGBA8 r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
    partial_store_span(dest, r, span);
  }
}

template <bool COMPOSITE>
static void linear_row_blit(uint8_t* dest, int span, const vec2_scalar& srcUV,
                            float srcDU, sampler2D sampler) {
  vec2 uv = init_interp(srcUV, vec2_scalar(srcDU, 0.0f));
  for (; span >= 4; span -= 4) {
    auto srcpx = textureLinearPackedR8(sampler, ivec2(uv));
    unaligned_store(dest, srcpx);
    dest += 4;
    uv.x += 4 * srcDU;
  }
  if (span > 0) {
    auto srcpx = textureLinearPackedR8(sampler, ivec2(uv));
    partial_store_span(dest, srcpx, span);
  }
}

template <bool COMPOSITE>
static void linear_row_blit(uint16_t* dest, int span, const vec2_scalar& srcUV,
                            float srcDU, sampler2D sampler) {
  vec2 uv = init_interp(srcUV, vec2_scalar(srcDU, 0.0f));
  for (; span >= 4; span -= 4) {
    auto srcpx = textureLinearPackedRG8(sampler, ivec2(uv));
    unaligned_store(dest, srcpx);
    dest += 4;
    uv.x += 4 * srcDU;
  }
  if (span > 0) {
    auto srcpx = textureLinearPackedRG8(sampler, ivec2(uv));
    partial_store_span(dest, srcpx, span);
  }
}

template <bool COMPOSITE = false>
static NO_INLINE void linear_blit(Texture& srctex, const IntRect& srcReq,
                                  Texture& dsttex, const IntRect& dstReq,
                                  bool invertX, bool invertY,
                                  const IntRect& clipRect) {
  assert(srctex.internal_format == GL_RGBA8 ||
         srctex.internal_format == GL_R8 || srctex.internal_format == GL_RG8);
  assert(!COMPOSITE || (srctex.internal_format == GL_RGBA8 &&
                        dsttex.internal_format == GL_RGBA8));
  // Compute valid dest bounds
  IntRect dstBounds = dsttex.sample_bounds(dstReq);
  dstBounds.intersect(clipRect);
  // Check if sampling bounds are empty
  if (dstBounds.is_empty()) {
    return;
  }
  // Initialize sampler for source texture
  sampler2D_impl sampler;
  init_sampler(&sampler, srctex);
  sampler.filter = TextureFilter::LINEAR;
  // Compute source UVs
  vec2_scalar srcUV(srcReq.x0, srcReq.y0);
  vec2_scalar srcDUV(float(srcReq.width()) / dstReq.width(),
                     float(srcReq.height()) / dstReq.height());
  if (invertX) {
    // Advance to the end of the row and flip the step.
    srcUV.x += srcReq.width();
    srcDUV.x = -srcDUV.x;
  }
  // Inverted Y must step downward along source rows
  if (invertY) {
    srcUV.y += srcReq.height();
    srcDUV.y = -srcDUV.y;
  }
  // Skip to clamped source start
  srcUV += srcDUV * (vec2_scalar(dstBounds.x0, dstBounds.y0) + 0.5f);
  // Scale UVs by lerp precision
  srcUV = linearQuantize(srcUV, 128);
  srcDUV *= 128.0f;
  // Calculate dest pointer from clamped offsets
  int bpp = dsttex.bpp();
  int destStride = dsttex.stride();
  char* dest = dsttex.sample_ptr(dstReq, dstBounds);
  int span = dstBounds.width();
  for (int rows = dstBounds.height(); rows > 0; rows--) {
    switch (bpp) {
      case 1:
        linear_row_blit<COMPOSITE>((uint8_t*)dest, span, srcUV, srcDUV.x,
                                   &sampler);
        break;
      case 2:
        linear_row_blit<COMPOSITE>((uint16_t*)dest, span, srcUV, srcDUV.x,
                                   &sampler);
        break;
      case 4:
        linear_row_blit<COMPOSITE>((uint32_t*)dest, span, srcUV, srcDUV.x,
                                   &sampler);
        break;
      default:
        assert(false);
        break;
    }
    dest += destStride;
    srcUV.y += srcDUV.y;
  }
}

// Whether the blit format is renderable.
static inline bool is_renderable_format(GLenum format) {
  switch (format) {
    case GL_R8:
    case GL_RG8:
    case GL_RGBA8:
      return true;
    default:
      return false;
  }
}

extern "C" {

void BlitFramebuffer(GLint srcX0, GLint srcY0, GLint srcX1, GLint srcY1,
                     GLint dstX0, GLint dstY0, GLint dstX1, GLint dstY1,
                     GLbitfield mask, GLenum filter) {
  assert(mask == GL_COLOR_BUFFER_BIT);
  Framebuffer* srcfb = get_framebuffer(GL_READ_FRAMEBUFFER);
  if (!srcfb) return;
  Framebuffer* dstfb = get_framebuffer(GL_DRAW_FRAMEBUFFER);
  if (!dstfb) return;
  Texture& srctex = ctx->textures[srcfb->color_attachment];
  if (!srctex.buf) return;
  Texture& dsttex = ctx->textures[dstfb->color_attachment];
  if (!dsttex.buf) return;
  assert(!dsttex.locked);
  if (srctex.internal_format != dsttex.internal_format &&
      (!is_renderable_format(srctex.internal_format) ||
       !is_renderable_format(dsttex.internal_format))) {
    // If the internal formats don't match, then we may have to convert. Require
    // that the format is a simple renderable format to limit combinatoric
    // explosion for now.
    assert(false);
    return;
  }
  // Force flipped Y onto dest coordinates
  if (srcY1 < srcY0) {
    swap(srcY0, srcY1);
    swap(dstY0, dstY1);
  }
  bool invertY = dstY1 < dstY0;
  if (invertY) {
    swap(dstY0, dstY1);
  }
  IntRect srcReq = IntRect{srcX0, srcY0, srcX1, srcY1} - srctex.offset;
  IntRect dstReq = IntRect{dstX0, dstY0, dstX1, dstY1} - dsttex.offset;
  if (srcReq.is_empty() || dstReq.is_empty()) {
    return;
  }
  IntRect clipRect = {0, 0, dstReq.width(), dstReq.height()};
  prepare_texture(srctex);
  prepare_texture(dsttex, &dstReq);
  if (!srcReq.same_size(dstReq) && srctex.width >= 2 && filter == GL_LINEAR &&
      srctex.internal_format == dsttex.internal_format &&
      is_renderable_format(srctex.internal_format)) {
    linear_blit(srctex, srcReq, dsttex, dstReq, false, invertY, dstReq);
  } else {
    scale_blit(srctex, srcReq, dsttex, dstReq, invertY, clipRect);
  }
}

typedef Texture LockedTexture;

// Lock the given texture to prevent modification.
LockedTexture* LockTexture(GLuint texId) {
  Texture& tex = ctx->textures[texId];
  if (!tex.buf) {
    assert(tex.buf != nullptr);
    return nullptr;
  }
  if (__sync_fetch_and_add(&tex.locked, 1) == 0) {
    // If this is the first time locking the texture, flush any delayed clears.
    prepare_texture(tex);
  }
  return (LockedTexture*)&tex;
}

// Lock the given framebuffer's color attachment to prevent modification.
LockedTexture* LockFramebuffer(GLuint fboId) {
  Framebuffer& fb = ctx->framebuffers[fboId];
  // Only allow locking a framebuffer if it has a valid color attachment.
  if (!fb.color_attachment) {
    assert(fb.color_attachment != 0);
    return nullptr;
  }
  return LockTexture(fb.color_attachment);
}

// Reference an already locked resource
void LockResource(LockedTexture* resource) {
  if (!resource) {
    return;
  }
  __sync_fetch_and_add(&resource->locked, 1);
}

// Remove a lock on a texture that has been previously locked
void UnlockResource(LockedTexture* resource) {
  if (!resource) {
    return;
  }
  if (__sync_fetch_and_add(&resource->locked, -1) <= 0) {
    // The lock should always be non-zero before unlocking.
    assert(0);
  }
}

// Get the underlying buffer for a locked resource
void* GetResourceBuffer(LockedTexture* resource, int32_t* width,
                        int32_t* height, int32_t* stride) {
  *width = resource->width;
  *height = resource->height;
  *stride = resource->stride();
  return resource->buf;
}

// Extension for optimized compositing of textures or framebuffers that may be
// safely used across threads. The source and destination must be locked to
// ensure that they can be safely accessed while the SWGL context might be used
// by another thread. Band extents along the Y axis may be used to clip the
// destination rectangle without effecting the integer scaling ratios.
void Composite(LockedTexture* lockedDst, LockedTexture* lockedSrc, GLint srcX,
               GLint srcY, GLsizei srcWidth, GLsizei srcHeight, GLint dstX,
               GLint dstY, GLsizei dstWidth, GLsizei dstHeight,
               GLboolean opaque, GLboolean flipX, GLboolean flipY,
               GLenum filter, GLint clipX, GLint clipY, GLsizei clipWidth,
               GLsizei clipHeight) {
  if (!lockedDst || !lockedSrc) {
    return;
  }
  Texture& srctex = *lockedSrc;
  Texture& dsttex = *lockedDst;
  assert(srctex.bpp() == 4);
  assert(dsttex.bpp() == 4);

  IntRect srcReq =
      IntRect{srcX, srcY, srcX + srcWidth, srcY + srcHeight} - srctex.offset;
  IntRect dstReq =
      IntRect{dstX, dstY, dstX + dstWidth, dstY + dstHeight} - dsttex.offset;
  if (srcReq.is_empty() || dstReq.is_empty()) {
    return;
  }

  // Compute clip rect as relative to the dstReq, as that's the same coords
  // as used for the sampling bounds.
  IntRect clipRect = {clipX - dstX, clipY - dstY, clipX - dstX + clipWidth,
                      clipY - dstY + clipHeight};
  // Ensure we have rows of at least 2 pixels when using the linear filter to
  // avoid overreading the row. Force X flips onto the linear filter for now
  // until scale_blit supports it.
  bool useLinear =
      srctex.width >= 2 &&
      (flipX || (!srcReq.same_size(dstReq) && filter == GL_LINEAR));

  if (opaque) {
    if (useLinear) {
      linear_blit<false>(srctex, srcReq, dsttex, dstReq, flipX, flipY,
                         clipRect);
    } else {
      scale_blit<false>(srctex, srcReq, dsttex, dstReq, flipY, clipRect);
    }
  } else {
    if (useLinear) {
      linear_blit<true>(srctex, srcReq, dsttex, dstReq, flipX, flipY, clipRect);
    } else {
      scale_blit<true>(srctex, srcReq, dsttex, dstReq, flipY, clipRect);
    }
  }
}

}  // extern "C"

// Saturated add helper for YUV conversion. Supported platforms have intrinsics
// to do this natively, but support a slower generic fallback just in case.
static inline V8<int16_t> addsat(V8<int16_t> x, V8<int16_t> y) {
#if USE_SSE2
  return _mm_adds_epi16(x, y);
#elif USE_NEON
  return vqaddq_s16(x, y);
#else
  auto r = x + y;
  // An overflow occurred if the signs of both inputs x and y did not differ
  // but yet the sign of the result did differ.
  auto overflow = (~(x ^ y) & (r ^ x)) >> 15;
  // If there was an overflow, we need to choose the appropriate limit to clamp
  // to depending on whether or not the inputs are negative.
  auto limit = (x >> 15) ^ 0x7FFF;
  // If we didn't overflow, just use the result, and otherwise, use the limit.
  return (~overflow & r) | (overflow & limit);
#endif
}

// Interleave and packing helper for YUV conversion. During transform by the
// color matrix, the color components are de-interleaved as this format is
// usually what comes out of the planar YUV textures. The components thus need
// to be interleaved before finally getting packed to BGRA format. Alpha is
// forced to be opaque.
static inline PackedRGBA8 packYUV(V8<int16_t> gg, V8<int16_t> br) {
  return pack(bit_cast<WideRGBA8>(zip(br, gg))) |
         PackedRGBA8{0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255};
}

// clang-format off
// Supports YUV color matrixes of the form:
// [R]   [1.1643835616438356,  0.0,  rv ]   [Y -  16]
// [G] = [1.1643835616438358, -gu,  -gv ] x [U - 128]
// [B]   [1.1643835616438356,  bu,  0.0 ]   [V - 128]
// We must be able to multiply a YUV input by a matrix coefficient ranging as
// high as ~2.2 in the U/V cases, where U/V can be signed values between -128
// and 127. The largest fixed-point representation we can thus support without
// overflowing 16 bit integers leaves us 6 bits of fractional precision while
// also supporting a sign bit. The closest representation of the Y coefficient
// ~1.164 in this precision is 74.5/2^6 which is common to all color spaces
// we support. Conversions can still sometimes overflow the precision and
// require clamping back into range, so we use saturated additions to do this
// efficiently at no extra cost.
// clang-format on
struct YUVMatrix {
  // These constants are loaded off the "this" pointer via relative addressing
  // modes and should be about as quick to load as directly addressed SIMD
  // constant memory.

  V8<int16_t> br_uvCoeffs;  // biased by 6 bits [b_from_u, r_from_v, repeats]
  V8<int16_t> gg_uvCoeffs;  // biased by 6 bits [g_from_u, g_from_v, repeats]
  V8<uint16_t> yCoeffs;     // biased by 7 bits
  V8<int16_t> yBias;        // 0 or 16
  V8<int16_t> uvBias;       // 128
  V8<int16_t> br_yMask;

  // E.g. rec709-narrow:
  // [ 1.16,     0,  1.79, -0.97 ]
  // [ 1.16, -0.21, -0.53,  0.30 ]
  // [ 1.16,  2.11,     0, -1.13 ]
  // =
  // [ yScale,        0, r_from_v ]   ([Y ]              )
  // [ yScale, g_from_u, g_from_v ] x ([cb] - ycbcr_bias )
  // [ yScale, b_from_u,        0 ]   ([cr]              )
  static YUVMatrix From(const vec3_scalar& ycbcr_bias,
                        const mat3_scalar& rgb_from_debiased_ycbcr,
                        int rescale_factor = 0) {
    assert(ycbcr_bias.z == ycbcr_bias.y);

    const auto rgb_from_y = rgb_from_debiased_ycbcr[0].y;
    assert(rgb_from_debiased_ycbcr[0].x == rgb_from_debiased_ycbcr[0].z);

    int16_t br_from_y_mask = -1;
    if (rgb_from_debiased_ycbcr[0].x == 0.0) {
      // gbr-identity matrix?
      assert(rgb_from_debiased_ycbcr[0].x == 0);
      assert(rgb_from_debiased_ycbcr[0].y >= 1);
      assert(rgb_from_debiased_ycbcr[0].z == 0);

      assert(rgb_from_debiased_ycbcr[1].x == 0);
      assert(rgb_from_debiased_ycbcr[1].y == 0);
      assert(rgb_from_debiased_ycbcr[1].z >= 1);

      assert(rgb_from_debiased_ycbcr[2].x >= 1);
      assert(rgb_from_debiased_ycbcr[2].y == 0);
      assert(rgb_from_debiased_ycbcr[2].z == 0);

      assert(ycbcr_bias.x == 0);
      assert(ycbcr_bias.y == 0);
      assert(ycbcr_bias.z == 0);

      br_from_y_mask = 0;
    } else {
      assert(rgb_from_debiased_ycbcr[0].x == rgb_from_y);
    }

    assert(rgb_from_debiased_ycbcr[1].x == 0.0);
    const auto g_from_u = rgb_from_debiased_ycbcr[1].y;
    const auto b_from_u = rgb_from_debiased_ycbcr[1].z;

    const auto r_from_v = rgb_from_debiased_ycbcr[2].x;
    const auto g_from_v = rgb_from_debiased_ycbcr[2].y;
    assert(rgb_from_debiased_ycbcr[2].z == 0.0);

    return YUVMatrix({ycbcr_bias.x, ycbcr_bias.y}, rgb_from_y, br_from_y_mask,
                     r_from_v, g_from_u, g_from_v, b_from_u, rescale_factor);
  }

  // Convert matrix coefficients to fixed-point representation. If the matrix
  // has a rescaling applied to it, then we need to take care to undo the
  // scaling so that we can convert the coefficients to fixed-point range. The
  // bias still requires shifting to apply the rescaling. The rescaling will be
  // applied to the actual YCbCr sample data later by manually shifting it
  // before applying this matrix.
  YUVMatrix(vec2_scalar yuv_bias, double yCoeff, int16_t br_yMask_, double rv,
            double gu, double gv, double bu, int rescale_factor = 0)
      : br_uvCoeffs(zip(I16(int16_t(bu * (1 << (6 - rescale_factor)) + 0.5)),
                        I16(int16_t(rv * (1 << (6 - rescale_factor)) + 0.5)))),
        gg_uvCoeffs(
            zip(I16(-int16_t(-gu * (1 << (6 - rescale_factor)) +
                             0.5)),  // These are negative coeffs, so
                                     // round them away from zero
                I16(-int16_t(-gv * (1 << (6 - rescale_factor)) + 0.5)))),
        yCoeffs(uint16_t(yCoeff * (1 << (6 + 1 - rescale_factor)) + 0.5)),
        // We have a +0.5 fudge-factor for -ybias.
        // Without this, we get white=254 not 255.
        // This approximates rounding rather than truncation during `gg >>= 6`.
        yBias(int16_t(((yuv_bias.x * 255 * yCoeff) - 0.5) * (1 << 6))),
        uvBias(int16_t(yuv_bias.y * (255 << rescale_factor) + 0.5)),
        br_yMask(br_yMask_) {
    assert(yuv_bias.x >= 0);
    assert(yuv_bias.y >= 0);
    assert(yCoeff > 0);
    assert(br_yMask_ == 0 || br_yMask_ == -1);
    assert(bu > 0);
    assert(rv > 0);
    assert(gu <= 0);
    assert(gv <= 0);
    assert(rescale_factor <= 6);
  }

  ALWAYS_INLINE PackedRGBA8 convert(V8<int16_t> yy, V8<int16_t> uv) const {
    // We gave ourselves an extra bit (7 instead of 6) of bias to give us some
    // extra precision for the more-sensitive y scaling.
    // Note that we have to use an unsigned multiply with a 2x scale to
    // represent a fractional scale and to avoid shifting with the sign bit.

    // Note: if you subtract the bias before multiplication, we see more
    // underflows. This could be fixed by an unsigned subsat.
    yy = bit_cast<V8<int16_t>>((bit_cast<V8<uint16_t>>(yy) * yCoeffs) >> 1);
    yy -= yBias;

    // Compute [B] = [yCoeff*Y + bu*U +  0*V]
    //         [R]   [yCoeff*Y +  0*U + rv*V]
    uv -= uvBias;
    auto br = br_uvCoeffs * uv;
    br = addsat(yy & br_yMask, br);
    br >>= 6;

    // Compute G = yCoeff*Y + gu*U + gv*V
    // First calc [gu*U, gv*V, ...]:
    auto gg = gg_uvCoeffs * uv;
    // Then cross the streams to get `gu*U + gv*V`:
    gg = addsat(gg, bit_cast<V8<int16_t>>(bit_cast<V4<uint32_t>>(gg) >> 16));
    // Add the other parts:
    gg = addsat(yy, gg);  // This is the part that needs the most headroom
                          // usually. In particular, ycbcr(255,255,255) hugely
                          // saturates.
    gg >>= 6;

    // Interleave B/R and G values. Force alpha (high-gg half) to opaque.
    return packYUV(gg, br);
  }
};

// Helper function for textureLinearRowR8 that samples horizontal taps and
// combines them based on Y fraction with next row.
template <typename S>
static ALWAYS_INLINE V8<int16_t> linearRowTapsR8(S sampler, I32 ix,
                                                 int32_t offsety,
                                                 int32_t stridey,
                                                 int16_t fracy) {
  uint8_t* buf = (uint8_t*)sampler->buf + offsety;
  auto a0 = unaligned_load<V2<uint8_t>>(&buf[ix.x]);
  auto b0 = unaligned_load<V2<uint8_t>>(&buf[ix.y]);
  auto c0 = unaligned_load<V2<uint8_t>>(&buf[ix.z]);
  auto d0 = unaligned_load<V2<uint8_t>>(&buf[ix.w]);
  auto abcd0 = CONVERT(combine(a0, b0, c0, d0), V8<int16_t>);
  buf += stridey;
  auto a1 = unaligned_load<V2<uint8_t>>(&buf[ix.x]);
  auto b1 = unaligned_load<V2<uint8_t>>(&buf[ix.y]);
  auto c1 = unaligned_load<V2<uint8_t>>(&buf[ix.z]);
  auto d1 = unaligned_load<V2<uint8_t>>(&buf[ix.w]);
  auto abcd1 = CONVERT(combine(a1, b1, c1, d1), V8<int16_t>);
  abcd0 += ((abcd1 - abcd0) * fracy) >> 7;
  return abcd0;
}

// Optimized version of textureLinearPackedR8 for Y R8 texture. This assumes
// constant Y and returns a duplicate of the result interleaved with itself
// to aid in later YUV transformation.
template <typename S>
static inline V8<int16_t> textureLinearRowR8(S sampler, I32 ix, int32_t offsety,
                                             int32_t stridey, int16_t fracy) {
  assert(sampler->format == TextureFormat::R8);

  // Calculate X fraction and clamp X offset into range.
  I32 fracx = ix;
  ix >>= 7;
  fracx = ((fracx & (ix >= 0)) | (ix > int32_t(sampler->width) - 2)) & 0x7F;
  ix = clampCoord(ix, sampler->width - 1);

  // Load the sample taps and combine rows.
  auto abcd = linearRowTapsR8(sampler, ix, offsety, stridey, fracy);

  // Unzip the result and do final horizontal multiply-add base on X fraction.
  auto abcdl = SHUFFLE(abcd, abcd, 0, 0, 2, 2, 4, 4, 6, 6);
  auto abcdh = SHUFFLE(abcd, abcd, 1, 1, 3, 3, 5, 5, 7, 7);
  abcdl += ((abcdh - abcdl) * CONVERT(fracx, I16).xxyyzzww) >> 7;

  // The final result is the packed values interleaved with a duplicate of
  // themselves.
  return abcdl;
}

// Optimized version of textureLinearPackedR8 for paired U/V R8 textures.
// Since the two textures have the same dimensions and stride, the addressing
// math can be shared between both samplers. This also allows a coalesced
// multiply in the final stage by packing both U/V results into a single
// operation.
template <typename S>
static inline V8<int16_t> textureLinearRowPairedR8(S sampler, S sampler2,
                                                   I32 ix, int32_t offsety,
                                                   int32_t stridey,
                                                   int16_t fracy) {
  assert(sampler->format == TextureFormat::R8 &&
         sampler2->format == TextureFormat::R8);
  assert(sampler->width == sampler2->width &&
         sampler->height == sampler2->height);
  assert(sampler->stride == sampler2->stride);

  // Calculate X fraction and clamp X offset into range.
  I32 fracx = ix;
  ix >>= 7;
  fracx = ((fracx & (ix >= 0)) | (ix > int32_t(sampler->width) - 2)) & 0x7F;
  ix = clampCoord(ix, sampler->width - 1);

  // Load the sample taps for the first sampler and combine rows.
  auto abcd = linearRowTapsR8(sampler, ix, offsety, stridey, fracy);

  // Load the sample taps for the second sampler and combine rows.
  auto xyzw = linearRowTapsR8(sampler2, ix, offsety, stridey, fracy);

  // We are left with a result vector for each sampler with values for adjacent
  // pixels interleaved together in each. We need to unzip these values so that
  // we can do the final horizontal multiply-add based on the X fraction.
  auto abcdxyzwl = SHUFFLE(abcd, xyzw, 0, 8, 2, 10, 4, 12, 6, 14);
  auto abcdxyzwh = SHUFFLE(abcd, xyzw, 1, 9, 3, 11, 5, 13, 7, 15);
  abcdxyzwl += ((abcdxyzwh - abcdxyzwl) * CONVERT(fracx, I16).xxyyzzww) >> 7;

  // The final result is the packed values for the first sampler interleaved
  // with the packed values for the second sampler.
  return abcdxyzwl;
}

// Casting to int loses some precision while stepping that can offset the
// image, so shift the values by some extra bits of precision to minimize
// this. We support up to 16 bits of image size, 7 bits of quantization,
// and 1 bit for sign, which leaves 8 bits left for extra precision.
const int STEP_BITS = 8;

// Optimized version of textureLinearPackedR8 for Y R8 texture with
// half-resolution paired U/V R8 textures. This allows us to more efficiently
// pack YUV samples into vectors to substantially reduce math operations even
// further.
template <bool BLEND>
static inline void upscaleYUV42R8(uint32_t* dest, int span, uint8_t* yRow,
                                  I32 yU, int32_t yDU, int32_t yStrideV,
                                  int16_t yFracV, uint8_t* cRow1,
                                  uint8_t* cRow2, I32 cU, int32_t cDU,
                                  int32_t cStrideV, int16_t cFracV,
                                  const YUVMatrix& colorSpace) {
  // As much as possible try to utilize the fact that we're only using half
  // the UV samples to combine Y and UV samples into single vectors. Here we
  // need to initialize several useful vector quantities for stepping fractional
  // offsets. For the UV samples, we take the average of the first+second and
  // third+fourth samples in a chunk which conceptually correspond to offsets
  // 0.5 and 1.5 (in 0..2 range). This allows us to reconstruct intermediate
  // samples 0.25, 0.75, 1.25, and 1.75 later. X fraction is shifted over into
  // the top 7 bits of an unsigned short so that we can mask off the exact
  // fractional bits we need to blend merely by right shifting them into
  // position.
  cU = (cU.xzxz + cU.ywyw) >> 1;
  auto ycFracX = CONVERT(combine(yU, cU), V8<uint16_t>)
                 << (16 - (STEP_BITS + 7));
  auto ycFracDX = combine(I16(yDU), I16(cDU)) << (16 - (STEP_BITS + 7));
  auto ycFracV = combine(I16(yFracV), I16(cFracV));
  I32 yI = yU >> (STEP_BITS + 7);
  I32 cI = cU >> (STEP_BITS + 7);
  // Load initial combined YUV samples for each row and blend them.
  auto ycSrc0 =
      CONVERT(combine(unaligned_load<V4<uint8_t>>(&yRow[yI.x]),
                      combine(unaligned_load<V2<uint8_t>>(&cRow1[cI.x]),
                              unaligned_load<V2<uint8_t>>(&cRow2[cI.x]))),
              V8<int16_t>);
  auto ycSrc1 = CONVERT(
      combine(unaligned_load<V4<uint8_t>>(&yRow[yI.x + yStrideV]),
              combine(unaligned_load<V2<uint8_t>>(&cRow1[cI.x + cStrideV]),
                      unaligned_load<V2<uint8_t>>(&cRow2[cI.x + cStrideV]))),
      V8<int16_t>);
  auto ycSrc = ycSrc0 + (((ycSrc1 - ycSrc0) * ycFracV) >> 7);

  // Here we shift in results from the next sample while caching results from
  // the previous sample. This allows us to reduce the multiplications in the
  // inner loop down to only two since we just need to blend the new samples
  // horizontally and then vertically once each.
  for (uint32_t* end = dest + span; dest < end; dest += 4) {
    yU += yDU;
    I32 yIn = yU >> (STEP_BITS + 7);
    cU += cDU;
    I32 cIn = cU >> (STEP_BITS + 7);
    // Load combined YUV samples for the next chunk on each row and blend them.
    auto ycSrc0n =
        CONVERT(combine(unaligned_load<V4<uint8_t>>(&yRow[yIn.x]),
                        combine(unaligned_load<V2<uint8_t>>(&cRow1[cIn.x]),
                                unaligned_load<V2<uint8_t>>(&cRow2[cIn.x]))),
                V8<int16_t>);
    auto ycSrc1n = CONVERT(
        combine(unaligned_load<V4<uint8_t>>(&yRow[yIn.x + yStrideV]),
                combine(unaligned_load<V2<uint8_t>>(&cRow1[cIn.x + cStrideV]),
                        unaligned_load<V2<uint8_t>>(&cRow2[cIn.x + cStrideV]))),
        V8<int16_t>);
    auto ycSrcn = ycSrc0n + (((ycSrc1n - ycSrc0n) * ycFracV) >> 7);

    // The source samples for the chunk may not match the actual tap offsets.
    // Since we're upscaling, we know the tap offsets fall within all the
    // samples in a 4-wide chunk. Since we can't rely on PSHUFB or similar,
    // instead we do laborious shuffling here for the Y samples and then the UV
    // samples.
    auto yshuf = lowHalf(ycSrc);
    auto yshufn =
        SHUFFLE(yshuf, yIn.x == yI.w ? lowHalf(ycSrcn).yyyy : lowHalf(ycSrcn),
                1, 2, 3, 4);
    if (yI.y == yI.x) {
      yshuf = yshuf.xxyz;
      yshufn = yshufn.xxyz;
    }
    if (yI.z == yI.y) {
      yshuf = yshuf.xyyz;
      yshufn = yshufn.xyyz;
    }
    if (yI.w == yI.z) {
      yshuf = yshuf.xyzz;
      yshufn = yshufn.xyzz;
    }

    auto cshuf = highHalf(ycSrc);
    auto cshufn =
        SHUFFLE(cshuf, cIn.x == cI.y ? highHalf(ycSrcn).yyww : highHalf(ycSrcn),
                1, 4, 3, 6);
    if (cI.y == cI.x) {
      cshuf = cshuf.xxzz;
      cshufn = cshufn.xxzz;
    }

    // After shuffling, combine the Y and UV samples back into a single vector
    // for blending. Shift X fraction into position as unsigned to mask off top
    // bits and get rid of low bits to avoid multiplication overflow.
    auto yuvPx = combine(yshuf, cshuf);
    yuvPx += ((combine(yshufn, cshufn) - yuvPx) *
              bit_cast<V8<int16_t>>(ycFracX >> (16 - 7))) >>
             7;

    // Cache the new samples as the current samples on the next iteration.
    ycSrc = ycSrcn;
    ycFracX += ycFracDX;
    yI = yIn;
    cI = cIn;

    // De-interleave the Y and UV results. We need to average the UV results
    // to produce values for intermediate samples. Taps for UV were collected at
    // offsets 0.5 and 1.5, such that if we take a quarter of the difference
    // (1.5-0.5)/4, subtract it from even samples, and add it to odd samples,
    // we can estimate samples 0.25, 0.75, 1.25, and 1.75.
    auto yPx = SHUFFLE(yuvPx, yuvPx, 0, 0, 1, 1, 2, 2, 3, 3);
    auto uvPx = SHUFFLE(yuvPx, yuvPx, 4, 6, 4, 6, 5, 7, 5, 7) +
                ((SHUFFLE(yuvPx, yuvPx, 4, 6, 5, 7, 4, 6, 5, 7) -
                  SHUFFLE(yuvPx, yuvPx, 5, 7, 4, 6, 5, 7, 4, 6)) >>
                 2);

    commit_blend_span<BLEND>(dest, colorSpace.convert(yPx, uvPx));
  }
}

// This is the inner loop driver of CompositeYUV that processes an axis-aligned
// YUV span, dispatching based on appropriate format and scaling. This is also
// reused by blendYUV to accelerate some cases of texture sampling in the
// shader.
template <bool BLEND = false>
static void linear_row_yuv(uint32_t* dest, int span, sampler2DRect samplerY,
                           const vec2_scalar& srcUV, float srcDU,
                           sampler2DRect samplerU, sampler2DRect samplerV,
                           const vec2_scalar& chromaUV, float chromaDU,
                           int colorDepth, const YUVMatrix& colorSpace) {
  // Calculate varying and constant interp data for Y plane.
  I32 yU = cast(init_interp(srcUV.x, srcDU) * (1 << STEP_BITS));
  int32_t yV = int32_t(srcUV.y);

  // Calculate varying and constant interp data for chroma planes.
  I32 cU = cast(init_interp(chromaUV.x, chromaDU) * (1 << STEP_BITS));
  int32_t cV = int32_t(chromaUV.y);

  // We need to skip 4 pixels per chunk.
  int32_t yDU = int32_t((4 << STEP_BITS) * srcDU);
  int32_t cDU = int32_t((4 << STEP_BITS) * chromaDU);

  if (samplerY->width < 2 || samplerU->width < 2) {
    // If the source row has less than 2 pixels, it's not safe to use a linear
    // filter because it may overread the row. Just convert the single pixel
    // with nearest filtering and fill the row with it.
    Float yuvF = {texelFetch(samplerY, ivec2(srcUV)).x.x,
                  texelFetch(samplerU, ivec2(chromaUV)).x.x,
                  texelFetch(samplerV, ivec2(chromaUV)).x.x, 1.0f};
    // If this is an HDR LSB format, we need to renormalize the result.
    if (colorDepth > 8) {
      int rescaleFactor = 16 - colorDepth;
      yuvF *= float(1 << rescaleFactor);
    }
    I16 yuv = CONVERT(round_pixel(yuvF), I16);
    commit_solid_span<BLEND>(
        dest,
        unpack(colorSpace.convert(V8<int16_t>(yuv.x),
                                  zip(I16(yuv.y), I16(yuv.z)))),
        span);
  } else if (samplerY->format == TextureFormat::R16) {
    // Sample each YUV plane, rescale it to fit in low 8 bits of word, and
    // then transform them by the appropriate color space.
    assert(colorDepth > 8);
    // Need to right shift the sample by the amount of bits over 8 it
    // occupies. On output from textureLinearUnpackedR16, we have lost 1 bit
    // of precision at the low end already, hence 1 is subtracted from the
    // color depth.
    int rescaleBits = (colorDepth - 1) - 8;
    for (; span >= 4; span -= 4) {
      auto yPx =
          textureLinearUnpackedR16(samplerY, ivec2(yU >> STEP_BITS, yV)) >>
          rescaleBits;
      auto uPx =
          textureLinearUnpackedR16(samplerU, ivec2(cU >> STEP_BITS, cV)) >>
          rescaleBits;
      auto vPx =
          textureLinearUnpackedR16(samplerV, ivec2(cU >> STEP_BITS, cV)) >>
          rescaleBits;
      commit_blend_span<BLEND>(
          dest, colorSpace.convert(zip(yPx, yPx), zip(uPx, vPx)));
      dest += 4;
      yU += yDU;
      cU += cDU;
    }
    if (span > 0) {
      // Handle any remaining pixels...
      auto yPx =
          textureLinearUnpackedR16(samplerY, ivec2(yU >> STEP_BITS, yV)) >>
          rescaleBits;
      auto uPx =
          textureLinearUnpackedR16(samplerU, ivec2(cU >> STEP_BITS, cV)) >>
          rescaleBits;
      auto vPx =
          textureLinearUnpackedR16(samplerV, ivec2(cU >> STEP_BITS, cV)) >>
          rescaleBits;
      commit_blend_span<BLEND>(
          dest, colorSpace.convert(zip(yPx, yPx), zip(uPx, vPx)), span);
    }
  } else {
    assert(samplerY->format == TextureFormat::R8);
    assert(colorDepth == 8);

    // Calculate varying and constant interp data for Y plane.
    int16_t yFracV = yV & 0x7F;
    yV >>= 7;
    int32_t yOffsetV = clampCoord(yV, samplerY->height) * samplerY->stride;
    int32_t yStrideV =
        yV >= 0 && yV < int32_t(samplerY->height) - 1 ? samplerY->stride : 0;

    // Calculate varying and constant interp data for chroma planes.
    int16_t cFracV = cV & 0x7F;
    cV >>= 7;
    int32_t cOffsetV = clampCoord(cV, samplerU->height) * samplerU->stride;
    int32_t cStrideV =
        cV >= 0 && cV < int32_t(samplerU->height) - 1 ? samplerU->stride : 0;

    // If we're sampling the UV planes at half the resolution of the Y plane,
    // then try to use half resolution fast-path.
    if (yDU >= cDU && cDU > 0 && yDU <= (4 << (STEP_BITS + 7)) &&
        cDU <= (2 << (STEP_BITS + 7))) {
      // Ensure that samples don't fall outside of the valid bounds of each
      // planar texture. Step until the initial X coordinates are positive.
      for (; (yU.x < 0 || cU.x < 0) && span >= 4; span -= 4) {
        auto yPx = textureLinearRowR8(samplerY, yU >> STEP_BITS, yOffsetV,
                                      yStrideV, yFracV);
        auto uvPx = textureLinearRowPairedR8(
            samplerU, samplerV, cU >> STEP_BITS, cOffsetV, cStrideV, cFracV);
        commit_blend_span<BLEND>(dest, colorSpace.convert(yPx, uvPx));
        dest += 4;
        yU += yDU;
        cU += cDU;
      }
      // Calculate the number of aligned chunks that we can step inside the
      // bounds of each planar texture without overreading.
      int inside = min(
          min((((int(samplerY->width) - 4) << (STEP_BITS + 7)) - yU.x) / yDU,
              (((int(samplerU->width) - 4) << (STEP_BITS + 7)) - cU.x) / cDU) *
              4,
          span & ~3);
      if (inside > 0) {
        uint8_t* yRow = (uint8_t*)samplerY->buf + yOffsetV;
        uint8_t* cRow1 = (uint8_t*)samplerU->buf + cOffsetV;
        uint8_t* cRow2 = (uint8_t*)samplerV->buf + cOffsetV;
        upscaleYUV42R8<BLEND>(dest, inside, yRow, yU, yDU, yStrideV, yFracV,
                              cRow1, cRow2, cU, cDU, cStrideV, cFracV,
                              colorSpace);
        span -= inside;
        dest += inside;
        yU += (inside / 4) * yDU;
        cU += (inside / 4) * cDU;
      }
      // If there are any remaining chunks that weren't inside, handle them
      // below.
    }
    for (; span >= 4; span -= 4) {
      // Sample each YUV plane and then transform them by the appropriate
      // color space.
      auto yPx = textureLinearRowR8(samplerY, yU >> STEP_BITS, yOffsetV,
                                    yStrideV, yFracV);
      auto uvPx = textureLinearRowPairedR8(samplerU, samplerV, cU >> STEP_BITS,
                                           cOffsetV, cStrideV, cFracV);
      commit_blend_span<BLEND>(dest, colorSpace.convert(yPx, uvPx));
      dest += 4;
      yU += yDU;
      cU += cDU;
    }
    if (span > 0) {
      // Handle any remaining pixels...
      auto yPx = textureLinearRowR8(samplerY, yU >> STEP_BITS, yOffsetV,
                                    yStrideV, yFracV);
      auto uvPx = textureLinearRowPairedR8(samplerU, samplerV, cU >> STEP_BITS,
                                           cOffsetV, cStrideV, cFracV);
      commit_blend_span<BLEND>(dest, colorSpace.convert(yPx, uvPx), span);
    }
  }
}

static void linear_convert_yuv(Texture& ytex, Texture& utex, Texture& vtex,
                               const YUVMatrix& rgbFromYcbcr, int colorDepth,
                               const IntRect& srcReq, Texture& dsttex,
                               const IntRect& dstReq, bool invertX,
                               bool invertY, const IntRect& clipRect) {
  // Compute valid dest bounds
  IntRect dstBounds = dsttex.sample_bounds(dstReq);
  dstBounds.intersect(clipRect);
  // Check if sampling bounds are empty
  if (dstBounds.is_empty()) {
    return;
  }
  // Initialize samplers for source textures
  sampler2DRect_impl sampler[3];
  init_sampler(&sampler[0], ytex);
  init_sampler(&sampler[1], utex);
  init_sampler(&sampler[2], vtex);

  // Compute source UVs
  vec2_scalar srcUV(srcReq.x0, srcReq.y0);
  vec2_scalar srcDUV(float(srcReq.width()) / dstReq.width(),
                     float(srcReq.height()) / dstReq.height());
  if (invertX) {
    // Advance to the end of the row and flip the step.
    srcUV.x += srcReq.width();
    srcDUV.x = -srcDUV.x;
  }
  // Inverted Y must step downward along source rows
  if (invertY) {
    srcUV.y += srcReq.height();
    srcDUV.y = -srcDUV.y;
  }
  // Skip to clamped source start
  srcUV += srcDUV * (vec2_scalar(dstBounds.x0, dstBounds.y0) + 0.5f);
  // Calculate separate chroma UVs for chroma planes with different scale
  vec2_scalar chromaScale(float(utex.width) / ytex.width,
                          float(utex.height) / ytex.height);
  vec2_scalar chromaUV = srcUV * chromaScale;
  vec2_scalar chromaDUV = srcDUV * chromaScale;
  // Scale UVs by lerp precision. If the row has only 1 pixel, then don't
  // quantize so that we can use nearest filtering instead to avoid overreads.
  if (ytex.width >= 2 && utex.width >= 2) {
    srcUV = linearQuantize(srcUV, 128);
    srcDUV *= 128.0f;
    chromaUV = linearQuantize(chromaUV, 128);
    chromaDUV *= 128.0f;
  }
  // Calculate dest pointer from clamped offsets
  int destStride = dsttex.stride();
  char* dest = dsttex.sample_ptr(dstReq, dstBounds);
  int span = dstBounds.width();
  for (int rows = dstBounds.height(); rows > 0; rows--) {
    linear_row_yuv((uint32_t*)dest, span, &sampler[0], srcUV, srcDUV.x,
                   &sampler[1], &sampler[2], chromaUV, chromaDUV.x, colorDepth,
                   rgbFromYcbcr);
    dest += destStride;
    srcUV.y += srcDUV.y;
    chromaUV.y += chromaDUV.y;
  }
}

// -
// This section must match gfx/2d/Types.h

enum class YUVRangedColorSpace : uint8_t {
  BT601_Narrow = 0,
  BT601_Full,
  BT709_Narrow,
  BT709_Full,
  BT2020_Narrow,
  BT2020_Full,
  GbrIdentity,
};

// -
// This section must match yuv.glsl

vec4_scalar get_ycbcr_zeros_ones(const YUVRangedColorSpace color_space,
                                 const GLuint color_depth) {
  // For SWGL's 8bpc-only pipeline, our extra care here probably doesn't matter.
  // However, technically e.g. 10-bit achromatic zero for cb and cr is
  // (128 << 2) / ((1 << 10) - 1) = 512 / 1023, which != 128 / 255, and affects
  // our matrix values subtly. Maybe not enough to matter? But it's the most
  // correct thing to do.
  // Unlike the glsl version, our texture samples are u8([0,255]) not
  // u16([0,1023]) though.
  switch (color_space) {
    case YUVRangedColorSpace::BT601_Narrow:
    case YUVRangedColorSpace::BT709_Narrow:
    case YUVRangedColorSpace::BT2020_Narrow: {
      auto extra_bit_count = color_depth - 8;
      vec4_scalar zo = {
          float(16 << extra_bit_count),
          float(128 << extra_bit_count),
          float(235 << extra_bit_count),
          float(240 << extra_bit_count),
      };
      float all_bits = (1 << color_depth) - 1;
      zo /= all_bits;
      return zo;
    }

    case YUVRangedColorSpace::BT601_Full:
    case YUVRangedColorSpace::BT709_Full:
    case YUVRangedColorSpace::BT2020_Full: {
      const auto narrow =
          get_ycbcr_zeros_ones(YUVRangedColorSpace::BT601_Narrow, color_depth);
      return {0.0, narrow.y, 1.0, 1.0};
    }

    case YUVRangedColorSpace::GbrIdentity:
      break;
  }
  return {0.0, 0.0, 1.0, 1.0};
}

constexpr mat3_scalar RgbFromYuv_Rec601 = {
    {1.00000, 1.00000, 1.00000},
    {0.00000, -0.17207, 0.88600},
    {0.70100, -0.35707, 0.00000},
};
constexpr mat3_scalar RgbFromYuv_Rec709 = {
    {1.00000, 1.00000, 1.00000},
    {0.00000, -0.09366, 0.92780},
    {0.78740, -0.23406, 0.00000},
};
constexpr mat3_scalar RgbFromYuv_Rec2020 = {
    {1.00000, 1.00000, 1.00000},
    {0.00000, -0.08228, 0.94070},
    {0.73730, -0.28568, 0.00000},
};
constexpr mat3_scalar RgbFromYuv_GbrIdentity = {
    {0, 1, 0},
    {0, 0, 1},
    {1, 0, 0},
};

inline mat3_scalar get_rgb_from_yuv(const YUVRangedColorSpace color_space) {
  switch (color_space) {
    case YUVRangedColorSpace::BT601_Narrow:
    case YUVRangedColorSpace::BT601_Full:
      return RgbFromYuv_Rec601;
    case YUVRangedColorSpace::BT709_Narrow:
    case YUVRangedColorSpace::BT709_Full:
      return RgbFromYuv_Rec709;
    case YUVRangedColorSpace::BT2020_Narrow:
    case YUVRangedColorSpace::BT2020_Full:
      return RgbFromYuv_Rec2020;
    case YUVRangedColorSpace::GbrIdentity:
      break;
  }
  return RgbFromYuv_GbrIdentity;
}

struct YcbcrInfo final {
  vec3_scalar ycbcr_bias;
  mat3_scalar rgb_from_debiased_ycbcr;
};

inline YcbcrInfo get_ycbcr_info(const YUVRangedColorSpace color_space,
                                GLuint color_depth) {
  // SWGL always does 8bpc math, so don't scale the matrix for 10bpc!
  color_depth = 8;

  const auto zeros_ones = get_ycbcr_zeros_ones(color_space, color_depth);
  const auto zeros = vec2_scalar{zeros_ones.x, zeros_ones.y};
  const auto ones = vec2_scalar{zeros_ones.z, zeros_ones.w};
  const auto scale = 1.0f / (ones - zeros);

  const auto rgb_from_yuv = get_rgb_from_yuv(color_space);
  const mat3_scalar yuv_from_debiased_ycbcr = {
      {scale.x, 0, 0},
      {0, scale.y, 0},
      {0, 0, scale.y},
  };

  YcbcrInfo ret;
  ret.ycbcr_bias = {zeros.x, zeros.y, zeros.y};
  ret.rgb_from_debiased_ycbcr = rgb_from_yuv * yuv_from_debiased_ycbcr;
  return ret;
}

// -

extern "C" {

// Extension for compositing a YUV surface represented by separate YUV planes
// to a BGRA destination. The supplied color space is used to determine the
// transform from YUV to BGRA after sampling.
void CompositeYUV(LockedTexture* lockedDst, LockedTexture* lockedY,
                  LockedTexture* lockedU, LockedTexture* lockedV,
                  YUVRangedColorSpace colorSpace, GLuint colorDepth, GLint srcX,
                  GLint srcY, GLsizei srcWidth, GLsizei srcHeight, GLint dstX,
                  GLint dstY, GLsizei dstWidth, GLsizei dstHeight,
                  GLboolean flipX, GLboolean flipY, GLint clipX, GLint clipY,
                  GLsizei clipWidth, GLsizei clipHeight) {
  if (!lockedDst || !lockedY || !lockedU || !lockedV) {
    return;
  }
  if (colorSpace > YUVRangedColorSpace::GbrIdentity) {
    assert(false);
    return;
  }
  const auto ycbcrInfo = get_ycbcr_info(colorSpace, colorDepth);
  const auto rgbFromYcbcr =
      YUVMatrix::From(ycbcrInfo.ycbcr_bias, ycbcrInfo.rgb_from_debiased_ycbcr);

  Texture& ytex = *lockedY;
  Texture& utex = *lockedU;
  Texture& vtex = *lockedV;
  Texture& dsttex = *lockedDst;
  // All YUV planes must currently be represented by R8 or R16 textures.
  // The chroma (U/V) planes must have matching dimensions.
  assert(ytex.bpp() == utex.bpp() && ytex.bpp() == vtex.bpp());
  assert((ytex.bpp() == 1 && colorDepth == 8) ||
         (ytex.bpp() == 2 && colorDepth > 8));
  // assert(ytex.width == utex.width && ytex.height == utex.height);
  assert(utex.width == vtex.width && utex.height == vtex.height);
  assert(ytex.offset == utex.offset && ytex.offset == vtex.offset);
  assert(dsttex.bpp() == 4);

  IntRect srcReq =
      IntRect{srcX, srcY, srcX + srcWidth, srcY + srcHeight} - ytex.offset;
  IntRect dstReq =
      IntRect{dstX, dstY, dstX + dstWidth, dstY + dstHeight} - dsttex.offset;
  if (srcReq.is_empty() || dstReq.is_empty()) {
    return;
  }

  // Compute clip rect as relative to the dstReq, as that's the same coords
  // as used for the sampling bounds.
  IntRect clipRect = {clipX - dstX, clipY - dstY, clipX - dstX + clipWidth,
                      clipY - dstY + clipHeight};
  // For now, always use a linear filter path that would be required for
  // scaling. Further fast-paths for non-scaled video might be desirable in the
  // future.
  linear_convert_yuv(ytex, utex, vtex, rgbFromYcbcr, colorDepth, srcReq, dsttex,
                     dstReq, flipX, flipY, clipRect);
}

}  // extern "C"